Pascal: What do we know? Discussion, Latest News & Updates: 1000M Series GPU's

HBM is a lot more power efficient than GDDR5.

 

Yeah, so I'm sort of on the right track when saying, "one step forward, two steps back."

 

With the clocks you have on mobile VRAM, the smaller 256bit memory bus they use, you save maybe around 15W using HBM over GDDR5

 

I think you guys might be a bit too optimistic about Pascal.
AMD worked with Hynix in developing HBM in the first place which is what gave them priority access to HBM 1 for a full YEAR before Nvidia could use it... as such, Nvidia wouldn't be able to use HBM until mid 2016 (if I'm not mistaken) ... however, the following news indicate that AMD also managed to secure priority access to HBM 2 as well, meaning that Nvidia might have to wait even longer and delay Pascal's release (that is of course if HBM supply remains relatively scarce):
http://wccftech.com/amd-working-entire-range-hbm-gpus-follow-fiji-fury-lineup/
http://hexus.net/tech/news/graphics/84662-amd-said-secured-priority-access-sk-hynixs-hbm2-chips/

 

NVIDIA won't delay Pascal again. It's already in production.

AMD prematurely released HBM to compete with NVIDIA's Maxwell. NVIDIA couldn't move on HBM until someone competed with Maxwell. It's illegal to keep moving forward without competition. It leads to monopolistic behavior, putting NVIDIA at an unfair advantage in the market.

These companies work on this stuff years in advance. NVIDIA is launching HBM2 before AMD in 2016. AMD is still re-branding their 300 series. We're not too optimistic. People <cough, Cloudfire> saying 990M will be more powerful than 980M SLI are likely too optimistic.

 

If Nvidia wants to use HBM 2 with Pascal, and if AMD secured priority access to HBM 2, then Nvidia will likely HAVE to delay (unless they intend on releasing Pascal with GDDR5).
Pascal itself may be in production, but they cannot finish it without HBM 2 (if they intend on using it).

How did AMD prematurely release HBM?
Nvidia can't use HBM 1 because that was something AMD secured for themselves for a full year because they worked with Hynix in developing HBM tech in the first place.
HBM 2 on the other hand isn't ready yet, and HBM as tech seems to be in relatively short supply, to which AMD seems to have priority access.
But, we cannot say anything with certainty until more time has passed.

[/quote]

I am aware that companies work years in advance on technologies, but your reasoning doesn't follow.

Hynix developed HBM technology by working with AMD.
AMD secured priority access to HBM 1 for a full year.
HBM 2 isn't ready yet (because its still being developed/finalized), and it would appear that Hynix is the supplier of this technology, to which AMD may have secured priority access due to their cooperation in developing the tech in the first place (which would be a good thing) - and furthermore, HBM 2 might seem to be in relatively short supply much like HBM 1.

Nvidia may have developed Pascal as a chip to work with HBM specs, but unless they get their hands on actual HBM 2 from Hynix, they cannot mass produce Pascal without it.
If I read and understood the articles correctly, neither AMD or NVidia can mass produce HBM 2 on their own.

 

They don't plan on producing Pascal without it. So, it must be in the works.

NVIDIA has said they've begun production of Pascal in May, preparing for a Q1 2016 launch. Did you read this article?

 

Exciting stuff with Pascal being on a smaller node too of course. I wonder how their new focus on compute will affect positively or negatively the gaming experience? They say in the article that 'unlike Maxwell, Nvidia has laid major focus on compute and GPGPU acceleration with Pascal' - I hope that this doesn't mean it's an optimised architecture for professional applications while not being able to tap gaming performance. Maxwell seems to be all about the gaming based on the fact that it's compute performance was rubbish, will be interesting to see if Pascal goes to far in the opposite direction?

 

I think NVIDIA is transitioning pretty hard into the realm of VR. That's where I believe the main focus with Pascal will be. It shouldn't affect gaming, per say. It should actually help with gaming, as gaming evolves. But it will be optimized for applications and other things in that field.

They're trying to dominate the market in GPU-Accelerated Deep Learning.

 

Sounds sensible. I certainly won't be upgrading until Pascal, maybe later, just depends on how my laptop runs games. Arkham Knight amazingly enough running well on it, despite the hate on the internet (bought a cheap key a week ago)!

 

Pascal should be ok for both gaming and compute. It has a new feature called "mixed precision".

If I understood well, it should mean with Pascal you should be able to get:

1x N of double-precision performance
2x N of single-precision performance
4x N of half-precision performance

depending on the workload needed.

Previous architectures instead had to have parts of silicon dedicated to different precisions (or simply paying the price of more expensive higher precision even if doing lower precision computations).

http://techreport.com/news/27978/nvidia-pascal-to-feature-mixed-precision-mode-up-to-32gb-of-ram

 

lol it's WCCF, and I think you linked the wrong article

Regardless Q1 2016 is being way too optimistic. I mean HBM2 has to enter volume production first, and then TSMC has to actually deliver on time, which if you know anything about their past history, is a pretty tall order.

There might be a limited production run in Q1, but I guarantee they'll all go to the compute segment, and consumer cards won't come until much later.

 

I've also seen this discussed and reported on ad nauseam too. AMD have first dibs on HBM2 for a limited time/yield due to their development with Hynix.
I don't see how Pascal can release first with HBM2 under these circumstances either although apparently ready.

 

OK, cool, so from reading that article & reading your post it seems that Pascal is pretty flexible and doesn't have to compromise on performance in different areas, whether that be gaming or professional applications - it's flexible enough to deal with all types of computation in an efficient manner.

 

From reading those articles sounds like AMD gets dibs on the first run on HBM2 inventory, didn't see anything about restrictions for Nvidia not being able to launch Pascal before AMD launches Arctic Isles.

Possible that Pascal has been ready for a bit just waiting on supply of HMB2 chips to become available to complete the piece?

 

Getting real mad at this stuff dude, been waiting for like 4 years to upgrade and decided Skylake Maxwell is the time and then I see articles that are like PASCELS GONNA BLOW YO MAXWELL UP IN DEM HOUSE. I'm just like =_=. Really?

 

And to think progress has slowed down significantly in recent years. 10-15 years ago if you were building a desktop you'd have a hard time keeping up with the blistering pace of hardware advancement.

 

It'll be worth waiting for. Pascal will be re-branded for two years. They re-branded the hell out of 28nm. Not much will change after this until 2019.

 

jinkizuite, Skylake with MXM should solve your problem.

 

Too bad I'm not a Clevo fan.

 

Always the same! You just got to make your judgements, not every Intel Gen is a big step, not every NVidia Gen (at least when judging by numbering alone - see 680M/780M/880M) is a big step, you wanna try & upgrade on the first big step of each to get maximum value I reckon - easier said than done, but for notebooks forget the CPU and concentrate on getting the right GPU (that's because Intel CPUs are all good, and also not improving fast, so makes no big difference).

 

680M, 780M, and 880M are bad examples because it's all built on the same backbone - re-brands.

You have to see 580M to 680M. This is the sort of jump we'll be seeing (again). And very likely, even better.

 

If you read my post again carefully you'll see that I'm saying that 680M/780M/880M weren't big jumps in performance. (Pay attention to the brackets, and the word ' not' that occurs shortly before the brackets.)

 

Eh I was gonna buy when the 980m came out but the products I had my eyes on had flaws with which I couldn't commit money immediately for. As time went on it was like Skylake + Windows 10 Maxwell seemed like the stars aligned for me and now this. When I look at how 980m only gets Witcher 3 to 30FPS its kinda a eh for me, yet I've really about had it with my current setup waiting for this holy grail pascal for another year is just such a hard carrot and stick bait.

 

My 980M gets 46-52 FPS in The Witcher 3. It was closer to the mid-30's when it first came out. Patches fixed that.

I wouldn't recommend buying anything now, though. Kind of late. Might as well wait until the end of this year for new stuff.

 

Witcher 3 is the new go to test for graphics power

 

The article in question mentions nothing about a Q1 2016 launch (unless you were referring to something else).
But here's an extract from the article:
"Also unlike Nvidia, AMD has a much more powerful incentive to launch its next generation of FinFET GPUs first. This is because the company has priority to HBM2 capacity – which is going to be limited initially – as a result of co-inventing the technology with Hynix. "

Based on this information (preliminary as it is), and if it remains accurate, Nvidia will not be able to release Pascal before AMD launches its own HBM2 gpu's.
Where is Nvidia supposed to get HBM2 to release Pascal in Q1 2016 BEFORE AMD if Hynix is the supplier, and they gave priority access to AMD?

 

are the graphics really that good maxed out? i've only seen videos and screen shots, but from what ive seen, crysis 3 looks much better maxed out

 

Probably just referring to Hairworks. Looks good, but halves performance.

 

i saw an nvidia slide that said arkham knight has '' faceworks'' yet i've seen no mention of it or comparison. what was that all about?

 

"HBM2 enables cards with 4 HBM 2.0 cards with 4GB per chip, or four HBM 2.0 cards with 8GB per chips results with 16GB and 32GB respectively. Pascal has power to do both, depending on the SKU. Pascal has 17 billion transistors and it will be significantly smaller silicon than the Maxwell 28nm based GM200. The second generation HBM (HBM 2.0) will enable 8Gb per DRAM die, 2Gbps speed per pin and 256 GB per second Bandwidth / stack." - Source

Just updating this thread. Seems to be a lot of repetition amongst sources.

 

TSMC starts mass production of 16nm chips! Confirmed production began at the end of last month.

"The 16nm FinFET+ node from TSMC is expected to deliver 2x the density, and 65%+ higher speed at 70%+ less power than the current 28HPM process. We will likely see an announcement by NVIDIA in Q1 2016 with Pascal-based products becoming available around Q2 2016." - Source

In other news...



Each HBM cube is connected to the GPU with a 1024bit wide memory bus. HBM modules actually operate at low frequencies compared to GDDR5, but thanks to the significantly wider memory interface they manage to be up to 9 times faster than standard GDDR5 memory modules.

 

I think I want to add that with a big 512bit memory bus operating at 1500-1750MHz and with 4GB GDDR5 vs 4GB HBM @ 500MHz the power saving is ~25W.

So take that "-70% less power than GDDR5" with some salt.
For mobile with much smaller 256bit memory bus and 1250MHz clocks the power savings will not be huge exactly.

 

Reducing the power consumption of the overall gpu by even just 25W might allow it
to be clocked higher.

 

That percentage off the GDDR5 would be about right for saving 25W.

 

I really liked all the fur on the animals in FC4 until I started skinning the larger animals, then I turned it off. Performance wasn't hit very much, but maybe that was just the specific title.

 

It does look quite amazing, graphically and visually. But I hate how it stomps on the GPU.

Pascal should be powerful enough to avoid such drastic FPS dips in the latest games, especially with NVLink, Mixed Precision, and DX12. SLI-scaling should also be near perfect (e.g. almost doubling performance with an additional card). 1080M SLI notebooks are going to be monsters.

 

Bad choice of words in some of that from me. Sorry about that.

My point was that 70% power reduction sounds super awesome while in reality its not much, and notebooks will even have less than that 25W reduction

 

Mobile HBM should also use ultra-wide memory bus too though? Otherwise it'd negate the whole point of using HBM.

 

It will be incidental to the desktop version.

 

8096bit makes no sense; surely you must mean 8192bit, right ?

 

Like 100% likely: 4096*2 = 8192...

 

On the high end they may just add extra power phases to reduce the quality of parts required.

 

Like this time?

 

My only question is, is how fast i can stuff this in a laptop. lol (i know its been answered many times.

im curious as do how the mobile version will stand up to the desktop version (apparently 980's didnt stray too far from ~30% gap)

and on a side note i have seen a lot of people say it wont be Q1 because of AMD's first pick of HBM but i hear samsung is jumping in on making it too.

Sent from my SGH-M919V using Tapatalk

 

It's likely NVIDIA will intentionally keep mobile GPU's performing below their desktop counterparts to avoid loss in sales of desktop components even though mobile GPU's may be able to perform equal to or better. This also allows them to re-brand mobile parts by adding a few cores that they could have added previously. It's typical behavior of pretty much any company developing internal components for computer technology - they invest in X and keep it on the market for as long as possible, making tiny tweaks and changes to re-brand it and sell as a new product. We haven't even seen full GM204 capabilities in mobile Maxwell, yet. This is exactly what I'm speaking of. They intentionally held back mobile Maxwell to squeeze as much money out of it as possible.

I'm sure mobile Pascal will be re-branded at least two times before Volta. I'm also going to assume Volta will be delayed 'till 2019.

 

yea, I could envision them doing that. though with any luck they will keep gimping to a minimum.

Too bad we dont live in a perfect world, eh?


Sent from my SGH-M919V using Tapatalk

 

_______________________________________________________

It really makes a lot of sense to see Samsung buy out and merge with AMD, and for so many reasons.
Sharing R&D platforms and Technologies is a no-brainer for why such a merger/acquisition should take place, if earlier rumors of a Samsung entrance into this foray are indeed true and forthcoming.

 

I agree. It's a good partnership, especially for AMD.

 

http://hexus.net/tech/news/ram/85751-samsung-start-hbm-mass-production-early-next-year/